On Mon, Nov 18, 2024, at 03:59, Maciej W. Rozycki wrote:
>
> to zero a 9-byte member at the byte offset of 1 of a quadword-aligned
> struct, happily clobbering a 1-byte member at the beginning of said struct
> if there is a concurrent or parallel write to that member in the middle of
> the unprotected RMW sequence.
>
> This patch series addresses these issues in the last two changes, having
> made generic test suite updates to improve coverage in the concerned area
> first and then having addressed a bunch of issues in the code affected I
> discovered in the course of this effort. There is a patch that includes
> pair of changes to the middle end (reload+LRA) required by this update as
> well, so it's not a purely backend-limited change, and hence no "Alpha:"
> prefix on the cover letter or the relevant patches.
I don't know enough about gcc internals to understand exactly
which problems you are addressing, but I think there are still
two issues that I identifier that remain unchanged here:
a) storing an '_Atomic' variable smaller than 8 bytes on non-bwx
targets should use ll/sc, but uses a plain rmw cycle.
b) Accessing a 'volatile' variable may have conflicting
requirements, and I'm not sure if either an ll/sc or a rmw
is actually correct since neither is free of side-effects.
Linux currently assumes that a store through a 'volatile
short *' pointer does not clobber adjacent data which
would need an atomic, but most CPUs (not sure about Alpha)
trap if you ever try an atomic operation on an MMIO
register address.
Arnd
